Installer polish
Keep improving Mac setup, guided install, global command opt-in, doctor output, and first-run troubleshooting.
Roadmap
Make local automation trustworthy enough to leave running.
SafeClaw’s roadmap is organized around install confidence, permission clarity, local model support, remote access, and contributor ergonomics.
Ollama flow
Add first-class local model setup: detect Ollama, offer model downloads, and write provider config without hand-editing env files.
Approval UX
Improve the approval path for terminal, chat, and WhatsApp so risky actions stay visible without becoming exhausting.
Tomorrow
Make the Mac chat interface feel fluid.
The Mac app chat should feel like a real assistant workspace, not a terminal log inside a window. The next pass is about speed, clarity, approvals, and smoother file/link workflows.
Streaming responses
Show partial assistant output as it arrives, with visible stop, retry, and failed states.
Recent sessions list
Add session history with rename, delete, export, timestamps, and quick switching.
Attachment preview drawer
Show dropped file name, size, preview mode, workspace warning, and send-as-reference versus include-content controls.
Approval cards
Render file writes, patches, shell commands, network fetches, and WhatsApp sends as clear allow/deny cards inside chat.
Message lifecycle
Track queued, running, needs approval, stopped, failed, and done states per message.
Clean result rendering
Separate assistant text from raw logs. Render command output, diffs, errors, and file references with dedicated UI blocks.
Starter actions
Add chips for Run doctor, Explain config, Check WhatsApp setup, Summarize folder, and Inspect dropped file.
Workspace awareness
Always show active workspace, model, permission profile, and approval mode above the composer.
Memory controls
Add remember, forget last message, search memory, and export session controls from the chat surface.
Open questions
Where contributors can have real impact.
The project is early enough that good feedback can still shape the safety model and product direction.
1
Permission profiles
Should SafeClaw use static profiles, per-tool approvals, per-session policy, or object-capability style grants?
2
Remote approvals
How should WhatsApp safely approve, deny, or queue actions when the user is away from the terminal?
3
Memory controls
What should users be able to inspect, redact, forget, export, and pin before long-running use feels comfortable?
Current focus
- safer defaults
- better install confidence
- local model support
- WhatsApp persistence
- clean docs
- package publishing
- higher test coverage
- contributor-friendly issuesCommunity
Curious builders, cautious automation, useful bug reports.
The best feedback is specific: what you tried, what broke, what felt unsafe, and what would make the assistant easier to trust.
Setup reports
Tell us where install instructions were confusing, especially on macOS and WhatsApp setup.
Safety critiques
Question the permission model. The project gets better when the boundaries get tested.
Tool ideas
Suggest small, inspectable tools that solve real workflows without widening access too casually.
Support
Support the next round of SafeClaw work.
Get the packaged Mac release or support hosting, testing, domain costs, and development time for safer local automation.